SenseTime has launched the Galaxy Project, a coordinated effort with nearly 20 partners to scale domestic AI chip infrastructure across China. In a keynote titled "Intelligent Transformation and Symbiosis," Yang Fan, the company's co-founder and president of its Large Device Business Group, described a closed loop connecting chip-level technology, ecosystem partnerships, and commercial deployment for domestically produced AI computing power. The initiative arrives as token demand climbs across enterprise deployments, industrial AI adoption catches up with consumer-facing use cases, and domestic chip commercialization reaches a point where intelligent computing centers built on Chinese silicon can be stood up at pace.
Alongside the Galaxy Project, SenseTime signed a space computing agreement with satellite manufacturer Guoxing Aerospace and struck a research partnership with five institutions — including the Shanghai Artificial Intelligence Laboratory — aimed at scientific computing applications. The company frames these moves as a response to three converging trends, though the window of opportunity depends heavily on performance figures that have not been independently verified. Enterprise buyers evaluating SenseTime's infrastructure should treat the company's projections as forecasts until quarterly results begin to land.
What's New / Specs
The Galaxy Project's stated ecosystem includes domestic chip vendors Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren Technology, component partner Xizhi Technology, and infrastructure firms Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology. SenseTime says the plan covers construction of one "token factory," five computing clusters at what it calls "10,000-calorie" scale, joint work across ten technology directions, and support for 200 AI startups.
- Token throughput claims: SenseTime says its large-scale device platform now processes an average of 2.42 trillion tokens daily, with a projection of 10 trillion tokens per day by Q4 2026 — a 25-fold increase.
- Hybrid inference performance: The company claims its heterogeneous hybrid inference technology delivers an 85–152 percent increase in Model FLOPs Utilisation on mainstream domestic chips, with inference cost-effectiveness at 1.25x that of Nvidia's H-series parts.
- Domestic homogeneous comparison: Compared with domestic homogeneous inference setups, SenseTime claims a 2.5x increase in token output at equivalent cost, which it says pushes optimised hybrid inference clusters past the minimum profitability threshold for domestic computing power.
- Full-stack adaptation layer: SenseTime says it has built a layer spanning models, frameworks, operators, toolchains, and hardware to let customers migrate workloads across domestic chip vendors without extensive rewrites.
- Applied examples: In an AI4S long-sequence protein prediction workload, fused operator optimisation reportedly cut overall prediction time by a factor of three. In AIGC video generation, the company claims a 93 percent multi-card parallel acceleration ratio for domestic chips running DiT models, alongside zero-cost migration for mainstream AI development tools.
- Energy efficiency metric: SenseTime introduced "Tokens Per Watt" as a replacement yardstick for AI data centre efficiency, alongside a Computing-Power Collaboration Agent handling resource scheduling, electricity price prediction, and energy storage optimisation across an eight-level data system with five decision chains.
- Energy claims: The company reports an 80 percent increase in token output per unit of electricity cost, average power prices 10 percent below comparable regional data centres, and 96 percent accuracy in computing load prediction.
None of these figures come with third-party benchmarking. The gap between a vendor's optimised test cluster and a customer's production environment — with its uneven data pipelines and delayed firmware updates — tends to be where such numbers soften. Domestic AI chips have historically struggled with a fragmented software stack: models trained for one architecture often require rework to run on another. SenseTime's adaptation layer aims to address that, but the claims read well in a sandbox test and matter far more once stress-tested against real customer pipelines running mixed hardware generations.
Beyond near-term infrastructure, SenseTime outlined work on optical computing for data centre efficiency, quantum computing applications in AI optimisation, and a space computing partnership with Guoxing Aerospace to build the SenseTime Space Computing Constellation. The plan calls for a first satellite launch in 2026, building toward thousands of computing satellites and computing capacity in the tens of thousands of petabytes by 2030. Yang argued the value extends past raw capability, framing space-based computing as a way to extend the reach of Chinese AI services into weak-network environments such as maritime operations and disaster response, and by extension to support China's AI exports internationally. That 2030 target sits five years out, and satellite computing deployments of this scale have no precedent to measure the timeline against.
Why It Matters
The Galaxy Project represents a systematic attempt to solve the multi-chip fragmentation that has hampered domestic AI infrastructure in China. By assembling a coalition spanning chip designers, component suppliers, and infrastructure integrators, SenseTime is trying to create a viable alternative to the Nvidia-dominated stack — not by matching peak performance on a single architecture, but by making heterogeneous domestic clusters economically workable at scale. The 2.5x token-output claim at equivalent cost, if reproducible in customer environments, would shift the economics of domestic computing power past a profitability threshold that has long been a barrier to adoption.
The physical infrastructure rollout underscores the ambition. SenseTime says its Shanghai facility runs the country's first data centre rated at "5A" intelligent computing level, handling over 20 trillion tokens daily across more than 20 industries. A Yancheng site has launched with an initial 3,000 petaflops of capacity focused on energy, manufacturing, and low-altitude economy applications. In Hong Kong, the company is building what it describes as the territory's largest domestic intelligent computing centre, targeting 40,000 petaflops by 2030. SenseTime also plans what it calls China's first overseas domestic computing cluster in Saudi Arabia, positioned as a full-stack domestic computing base for the Middle East. On the research side, the tie-up with the Shanghai AI Laboratory, Beijing Zhongguancun Academy, Shenzhen Hetao Academy, the Shanghai Algorithm Innovation Research Institute, and Shanghai Jiao Tong University's AI school aims to build a shared platform spanning compute, tooling, and model capability for life sciences, materials science, and manufacturing research.
Yang called AI for Science "a key lever for paradigm innovation in basic research," tying the initiative to China's broader "Artificial Intelligence+" policy push. The space computing bet — thousands of satellites delivering tens of thousands of petabytes by 2030 — is a long-horizon play with no direct precedent. Electricity price arbitrage and load forecasting accuracy tend to perform differently once a system runs through a full seasonal cycle with genuine demand volatility, rather than the conditions under which a vendor typically runs its pilot. The Tokens Per Watt metric and Computing-Power Collaboration Agent are claims worth watching over the next several quarters rather than accepting at face value.
Our Take
SenseTime's Galaxy Project is the most comprehensive domestic AI infrastructure play yet announced in China, and the partner roster reads like a who's who of the country's chip ecosystem. The heterogeneous hybrid inference approach is technically sound: different domestic architectures have complementary strengths, and a software layer that abstracts those differences could unlock utilisation rates that homogeneous clusters cannot achieve. However, the performance claims carry the usual vendor caveats. The 85–152 percent Model FLOPs Utilisation improvement, the 1.25x cost-effectiveness versus Nvidia H-series, and the 2.5x token-output gain over domestic homogeneous setups are all self-reported, measured on optimised test clusters that may not reflect the heterogeneity of customer workloads.
The adaptation layer is the critical piece. If SenseTime can genuinely deliver zero-cost migration across Cambricon, Muxi, Hygon, Ascend, Moore Threads, Sunrise, and Biren without extensive rewrites, that alone would be a significant advance — but the history of cross-vendor compatibility in AI accelerators is littered with abstraction layers that work for benchmark models and break on production pipelines. The AI4S protein prediction and AIGC video generation examples are encouraging, yet they represent narrow, well-understood workloads. The real test comes when customers bring mixed model families, dynamic batching requirements, and legacy framework dependencies.
The energy metrics and space computing roadmap are ambitious to the point of speculation. An 80 percent improvement in token output per electricity cost, 10 percent lower power prices, and 96 percent load prediction accuracy are extraordinary claims for a system that has not yet completed a full seasonal cycle. The 2030 satellite constellation target — thousands of satellites, tens of thousands of petabytes — has no precedent in any national program. The Saudi Arabia cluster is a concrete near-term milestone that will indicate whether the full-stack domestic offering can compete internationally. For now, the figure to track is the 10 trillion tokens per day forecast for Q4 2026. When that quarter closes, the reported number will tell us whether the Galaxy Project's closed loop is closing at the pace SenseTime claims.
FAQ
What is the SenseTime Galaxy Project?
The Galaxy Project is SenseTime's initiative to scale domestic AI chip infrastructure in China by coordinating nearly 20 partners across chip design, components, and infrastructure. It aims to build a "token factory," five large-scale computing clusters, joint R&D across ten technology directions, and support for 200 AI startups, all running on Chinese-made silicon.
What performance claims does SenseTime make for its hybrid inference technology?
SenseTime claims an 85–152 percent increase in Model FLOPs Utilisation on mainstream domestic chips, inference cost-effectiveness at 1.25x that of Nvidia H-series GPUs, and a 2.5x increase in token output versus domestic homogeneous inference setups at equivalent cost. These figures are self-reported and have not been independently benchmarked.
How does SenseTime address the fragmented domestic chip software stack?
The company says it has built a full-stack adaptation layer spanning models, frameworks, operators, toolchains, and hardware, designed to let customers migrate workloads across domestic chip vendors — including Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren — without extensive rewrites.
What is the Tokens Per Watt metric and Computing-Power Collaboration Agent?
Tokens Per Watt is SenseTime's proposed replacement yardstick for AI data centre efficiency. The Computing-Power Collaboration Agent handles resource scheduling, electricity price prediction, and energy storage optimisation across an eight-level data system with five decision chains. SenseTime claims an 80 percent increase in token output per electricity cost, 10 percent lower average power prices, and 96 percent load prediction accuracy.
What are the space computing and overseas expansion plans?
SenseTime and Guoxing Aerospace plan the SenseTime Space Computing Constellation, targeting a first satellite launch in 2026 and thousands of computing satellites with tens of thousands of petabytes capacity by 2030. On the ground, SenseTime is building a domestic computing cluster in Saudi Arabia as a full-stack base for the Middle East, alongside facilities in Shanghai, Yancheng, and Hong Kong targeting 40,000 petaflops by 2030.
